metadata
language:
- en
license: mit
pretty_name: Wiki-Simple-Cleaned
tags:
- wikipedia
- cleaned
- pretraining
- simple-english
task_categories:
- text-generation
Wiki-Simple-Cleaned
A cleaned version of Simple English Wikipedia.
Dataset Curation
- Source: Based on legacy-datasets/wikipedia.
- Noise Removal: Stripped References, See also, External links, and Sources.
- Formatting: Headers converted to sentences; newlines and multiple spaces removed.
- Special Tokens: Documents wrapped in
<|bos|>{Title}. {Content}<|eos|>.
Dataset Structure
- Format: Plain text (.txt) where each line is a single document.
- Content: Simple English, optimized for models under 500M parameters.
- Distribution: ~70% short, ~27% medium, and ~3% long-form articles.